A Statistical-Computational Framework for Privacy-Preserving Federated Reliability Inference
A Statistical-Computational Framework for Privacy-Preserving Federated Reliability Inference, Lefkoşa, Kıbrıs (Kktc), 22 - 24 Ekim 2026, ss.69, (Özet Bildiri)
- Yayın Türü: Bildiri / Özet Bildiri
- Basıldığı Şehir: Lefkoşa
- Basıldığı Ülke: Kıbrıs (Kktc)
- Sayfa Sayıları: ss.69
- Ankara Üniversitesi Adresli: Evet
Özet
Privacy-preserving reliability analysis often requires combining failure information held by multiple assets or organizations without centralizing raw data. This study develops a statistical-computational framework in which the communication structure is determined directly by the likelihood representation. Three cases are distinguished: one-shot aggregation based on parameter-independent summaries, iterative protected evaluation when local likelihood contributions depend on the current parameter, and local elimination of nuisance parameters before assigning the communication regime. The framework is evaluated using two engineering reliability benchmarks. For the 13-aircraft Proschan data, exponential estimation and fleet-homogeneity testing are recovered from a single protected summary. The two-parameter Gamma lifetime model is also one-shot because its pooled likelihood is completely determined by the sample size, the sum of lifetimes, and the sum of log-lifetimes. Weibull inference requires parameter-dependent evaluations; a safeguarded profile-Newton procedure reaches the centralized maximum-likelihood estimate in five protected evaluations from the reference start and converges to the same solution from all nine examined positive starting values. For the failure-truncated power-law process, conditioning removes aircraft-specific scale parameters and yields exact finitesample common-trend inference, with β̂ = 1.13537 and p = 0.07882. The Farley-1 nuclear-pump application further illustrates one-shot conditional Poisson inference and iterative Gamma-Poisson random-effects estimation. The intermittent-to-continuous failure-rate ratio is 6.57. Across the studied models, protected estimates match their centralized counterparts to numerical tolerance. The results show that federated reliability inference can be made more communicationefficient by exploiting statistical representation before introducing iterative optimization.